VRAM is the first constraint, but not the whole server. Start with model size, precision, concurrency, context length, and dataset size, then consider GPU count, CPU, system RAM, NVMe, bandwidth, and runtime. Training, inference, rendering, and research emphasize different combinations of bandwidth, interconnect, disk, and system memory.
Write multi-GPU capacity as “VRAM per GPU × card count = aggregate VRAM.” Two 32 GB GPUs provide 64 GB aggregate VRAM, but software parallelism and model sharding determine whether the workload can use it effectively; it is not automatically one contiguous 64 GB pool.
CPU, RAM, and NVMe should follow preprocessing, compilation, caching, checkpoint, and multi-user needs. The configurator submits these requirements with the GPU, term, and region; the confirmed product facts, inventory, and order are the final source of truth.